Research And Implementation Of Driver Distraction Behavior Detection Based On Video Sequence

Road traffic safety has always been a hot issue that everyone is concerned about,and distracted driving by drivers is one of the important causes of road traffic accidents.With the diversified development of in-vehicle entertainment systems and the rapid advancement of automatic driving technology,the driver’s attention and its easy dispersal are often prone to serious fatalities and injuries.Driver distraction detection can detect driver distraction behavior and identify the cause of distraction,and then warn the driver not to be distracted or take preventive measures.On the other hand,in the case of automatic driving,it helps to determine whether the driver has the ability to take over the vehicle.The car can stop the distracting behavior by warning the driver and concentrate on driving the vehicle to reduce the occurrence of accidents.This makes the driver distraction detection system an important system component of vehicle intelligence.This paper proposes a multi-information fusion driver distraction detection method based on graph convolution.In this paper,a driver distracted driving video clip dataset is produced,and a driver distraction detection hardware platform is built in the Jetson TX2 embedded system,and the accuracy,real-time and robustness of the algorithm are verified.The innovation and work of this article are as follows:(1)According to the cab environment and the driver’s attitude characteristics,a time-space map of the driver’s attitude is designed.Using Openpose to extract the key points of the human body,and according to the characteristics of the driver’s movement concentrated on the hand and head,the driver’s posture estimation map was improved.In view of the current lack of driver distraction video data sets,the driver distracted driving video clip data set was produced from the aspects of shooting angle,driver body size,lighting,interference,etc.,and the driver distracted data set was converted into driver Attitude space-time map data set.(2)Aiming at the non-Euclidean spatial data structure of the driver’s posture estimation space-time graph,based on the graph convolution network and fusing the skeleton and key object information,a driver distraction behavior detection algorithm is proposed.Aiming at the requirement of high robustness for driver distraction detection,the algorithm in this paper extracts the features of driver skeleton diagram in time and space based on graph convolution network,and uses the fully connected layer to discriminate the driver’s distraction behavior.By combining the co-occurrence features of key objects and human joints,a highly robust driver distraction behavior detection model can be obtained in real time.(3)Aiming at the problem that the algorithm in this paper cannot run in real time in the embedded system,Mobile Net and hollow convolution are used to lighten the driver’s distraction detection algorithm,reduce the amount of calculation,and speed up the calculation speed.At the same time,a driver distraction detection platform was built on the Jetson TX2 developer kit,and the lightweight algorithm was verified on the platform.